From one of our Search devs:
As of today it’s using a combination of llama and qwen, always open models we can run in-house to guarantee privacy. Most of the work is not on the model but on the selection of what data is chosen as context. We switch models very, very often So this will be outdated soon.